Can LLM-Augmented Autonomous Agents Cooperate? An Evaluation of Their Cooperative Capabilities Through Melting Pot

As artificial intelligence continues to advance, a key aspect of this progression is the development of Large Language Models (LLMs) and their capacity to enhance multi-agent artificial intelligence systems. This paper investigates the cooperative cabilities of Large Language Model-augmented Autonomous Agents (LAAs) using the well-known Melting Pot environments along with reference models such as GPT-4o, GPT-4o mini, GPT-4, and GPT-3.5. Preliminary results suggest that comprehensive LAA architectures significantly improve performance in Melting Pot’s multi-agent scenarios, outperforming reinforcement learning baselines in two out of three scenarios and surpassing a simpler Chain-of-Thought agent architecture in all scenarios. The Melting Pot environments are specifically designed to assess multi-agent systems, requiring agents to exhibit cooperative behavior to solve conditions effectively.